Vous êtes sur la page 1sur 26

Adjusted Survival Curves

Terry M Therneau, Cynthia S Crowson, Elizabeth J Atkinson


Jan 2015

1 Introduction
Suppose we want to investigate to what extent some factor influences survival, as an example
we might compare the experience of diabetic patients who are using metformin versus those on
injected insulin as their primary treatment modality. There is some evidence that metformin
has a positive influence, particularly in cancers, but the ascertainment is confounded by the fact
that it is a first line therapy: the patients on metformin will on average be younger and have
had a diabetes diagnosis for a shorter amount of time than those using insulin. “Young people
live longer” is not a particularly novel observation.
The ideal way to test this is with a controlled clinical trial. This is of course not always
possible, and assessments using available data that includes and adjusts for such confounders is
also needed. There is extensive literature — and debate — on this topic in the areas of modeling
and testing. The subtopic of how to create honest survival curve estimates in the presence of
confounders is less well known, and is the focus of this note.
Assume that we have an effect of interest, treatment say, and a set of possible confounding
variables. Creation a pair of adjusted survival curves has two parts: definition of a reference
population for the confounders, and then the computation of estimated curves for that popula-
tion. There are important choices in both steps. The first, definition of a target, is often not
explicitly stated but can be critical. If an outcome differs with age, myocardial infarction say,
and two treatments also had age dependent efficacy, then the comparison will depend greatly on
whether we are talking about a population of young, middle aged, or older subjects.
The computational step has two main approaches. The first, sometimes known as marginal
analysis, first reweights the data such that each subgroup’s weighted distribution matches that
of our population target. An immediate consequence is that all subgroups will be balanced
with respect to the confounding variables. We can then proceed with a simple analysis of
survival using the reformulated data, ignoring the confounders. The second approach seeks to
understand and model the effect of each confounder, with this we can then correct for them. From
a comprehensive overall model we can obtain predicted survival curves for any configuration of
variables, and from these get predicted overall curves for the reference population. This is often
called the conditional approach since we are using conditional survival curves given covariates x.
A third but more minor choice is division of the covariates x into effects of interest vs. con-
founders. For instance, we might want to see separate curves for two treatments, each adjusted
for age and sex. The reference population will describe the age and sex distribution. For simplic-
ity we will use x to describe all the confounding variables and use c for the control variable(s),

1
1.0

FLC < 3.38


0.8
0.6

3.38 − 4.71
Survival

0.4

FLC > 4.71


0.2
0.0

0 2 4 6 8 10 12 14

Years from Sample

Figure 1: Survival of 7874 residents of Olmsted County, broken into three cohorts based on FLC
value.

e.g. treatment. The set c might be empty, producing a single overall curve, but this is the
uncommon case. As shown below, our two methods differ essentially in the order in which the
two necessary operations are done, balancing and survival curve creation.
Marginal: balance data on x −→ form survival curves for each c
Conditional: predicted curves for {x, c} subset −→ average the predictions for each c
We can think of them as “balance and then model” versus “model then balance”. An analysis
might use a combinations of these, of course, balancing on some factors and modeling others.
All analyses are marginal analyses with respect to important predictors that are unknown to us,
although in that case we have no assurance of balance on those factors.

2 Free Light Chain


Our example data set for this comparison uses a particular assay of plasma immunoglobulins and
is based on work of Dr Angela Dispenzieri and her colleagues at the Mayo Clinic [2]. In brief:
plasma cells (PC) are responsible for the production of immunoglobulins, but PC comprise only
a small portion (< 1%) of the total blood and marrow hematapoetic cell population in normal

2
50–59 60–69 70–79 80+
FLC < 3.38 2585 (47) 1690 (30) 969 (17) 314 ( 6)
FLC 3.38–4.71 442 (29) 446 (29) 423 (28) 215 (14)
FLC > 4.71 121 (16) 187 (25) 224 (30) 224 (30)

Table 1: Comparison of the age distributions (percents) for each of the three groups.

patients. The normal human repertoire is estimated to contain over 108 unique immunoglobulins,
conferring a broad range of immune protection. In multiple myeloma, the most common form
of plasma cell malignancy, almost all of the circulating antigen will be identical, the product
of a single malignant clone. An electrophoresis examination of circulating immunoglobulins will
exhibit a “spike” corresponding to this unique molecule. This anomaly is used both as a diagnostic
method and in monitoring the course of the disease under treatment.
The presence of a similar, albeit much smaller, spike in normal patients has been a long
term research interest of the Mayo Clinic hematology research group [5]. In 1995 Dr Robert
Kyle undertook a population based study of this, and collected serum samples on 19,261 of the
24,539 residents of Olmsted County, Minnesota, aged 50 years or more [4]. In 2010 Dr. Angela
Dispenzieri assayed a sub fraction of the immunoglobulins, the free light chain (FLC), on 15,748
of these subjects who had sufficient remaining sera from the original sample collection. All
studies took place under the oversight of the appropriate Institutional Review Boards, which
ensure rigorous safety and ethical standards in research.
A subset of the Dispenzieri study is available in the survival package as data set flchain.
Because the original study assayed nearly the entire population, there is concern that some
portions of the anonymized data could be linked to actual subjects by a diligent searcher, and
so only a subset of the study has been made available as a measure to strengthen anonymity.
It was randomly selected from the whole within sex and age group strata so as to preserve the
age/sex structure. The data set contains 3 subjects whose blood sample was obtained on the
day of their death. It is rather odd to think of a sample obtained on the final day as “predicting”
death, or indeed for any results obtained during a patient’s final mortality cascade. There are
also a few patients with no follow-up beyond the clinic visit at which the assay occurred. We
have chosen in this analysis to exclude the handful of subjects with less than 7 days of follow-up,
leaving 7840 observations.
Figure 1 shows the survival curves for three subgroups of the patients: those whose total free
light chain (FLC) is in the upper 10% of all values found in the full study, those in the 70–89th
percentile, and the remainder. There is a clear survival effect. Average free light chain amounts
rise with age, however, at least in part because it is eliminated through the kidneys and renal
function declines with age. Table 1 shows the age distribution for each of the three groups. In
the highest decile of FLC (group 3) over half the subjects are age 70 or older compared to only
23% in those below the 70th percentile. How much of the survival difference is truly associated
with FLC and how much is simply an artifact of age? (The cut points are arbitrary, but we have
chosen to mimic the original study and retain them. Division into three groups is a convenient
number to illustrate the methods in this vignette, but we do not make any claim that such a
categorization is optimal or even sensible statistical practice.) The R code for figure 1 is shown
below.
> fdata <- flchain[flchain$futime >=7,]

3
> fdata$age2 <- cut(fdata$age, c(0,54, 59,64, 69,74,79, 89, 110),
labels = c(paste(c(50,55,60,65,70,75,80),
c(54,59,64,69,74,79,89), sep='-'), "90+"))
> fdata$group <- factor(1+ 1*(fdata$flc.grp >7) + 1*(fdata$flc.grp >9),
levels=1:3,
labels=c("FLC < 3.38", "3.38 - 4.71", "FLC > 4.71"))
> sfit1 <- survfit(Surv(futime, death) ~ group, fdata)
> plot(sfit1, mark.time=F, col=c(1,2,4), lty=1, lwd=2,
xscale=365.25, xlab="Years from Sample",
ylab="Survival")
> text(c(11.1, 10.5, 7.5)*365.25, c(.88, .57, .4),
c("FLC < 3.38", "3.38 - 4.71", "FLC > 4.71"), col=c(1,2,4))

3 Reference populations
There are a few populations that are commonly used as the reference group.
1. Empirical. The overall distribution of confounders x in the data set as a whole. For this
study we would use the observed age/sex distribution, ignoring FLC group. This is also
called the “sample” or “data” distribution.
2. External reference. The distribution from some external study or standard.
3. Internal reference. A particular subset of the data is chosen as the reference, and other
subsets are then aligned with it.
Method 2 is common in epidemiology, using a reference population based on a large external
population such as the age/sex distribution of the 2000 United States census. Method 3 most
often arises in the case-control setting, where one group is small and precious (a rare disease
say) and the other group (the controls) from which we can sample is much larger. In each case
the final result of the computation can be thought of as the expected answer we “would obtain”
in a study that was perfectly balanced with respect to the list of confounders x. Population 1 is
the most frequent.

4 Marginal approach
4.1 Selection
One approach for balancing is to select a subset of the data such that its distribution matches
the referent for each level of c, i.e., for each curve that we wish to obtain. As an example we
take a case-control like approach to the FLC data, with FLC high as the “cases” since it is the
smallest group. Table 2 shows a detailed distribution of the data with respect to age and sex.
The balanced subset has all 32 females aged 50–54 from the high FLC group, a random sample of
32 out of the 738 females in the age 50–54 low FLC group, and 32 out of 110 for the middle FLC.
Continue this for all age/sex subsets. We cannot quite compute a true case-control estimate for
this data since there are not enough “controls” in the female 90+ category to be able to select

4
Females
Age
FLC group 50–54 55–59 60–64 65–69 70–74 75–79 80–89 90+
Low 738 638 490 442 376 254 224 24
Med 110 94 88 104 116 108 125 21
High 32 30 45 43 48 46 103 32

Males
Low 658 553 427 331 227 112 64 2
Med 111 127 124 130 110 90 63 6
High 26 33 39 61 66 65 74 15

Table 2: Detailed age and sex distribution for the study population

one unique control for each case, and likewise in the male 80-89 and 90+ age groups. To get
around this we will sample with replacement in these strata.

> temp <- with(fdata, table(group, age2, sex))


> dd <- dim(temp)
> # Select subjects
> set.seed(1978)
> select <- array(vector('list', length=prod(dd)), dim=dd)
> for (j in 1:dd[2]) {
for (k in 1:dd[3]) {
n <- temp[3,j,k] # how many to select
for (i in 1:2) {
indx <- which(as.numeric(fdata$group)==i &
as.numeric(fdata$age2) ==j &
as.numeric(fdata$sex) ==k)
select[i,j,k] <- list(sample(indx, n, replace=(n> temp[i,j,k])))
}
indx <- which(as.numeric(fdata$group)==3 &
as.numeric(fdata$age2) ==j &
as.numeric(fdata$sex) ==k)
select[3,j,k] <- list(indx) #keep all the group 3 = high
}
}
> data2 <- fdata[unlist(select),]
> sfit2 <- survfit(Surv(futime, death) ~ group, data2)
> plot(sfit2,col=c(1,2,4), lty=1, lwd=2,
xscale=365.25, xlab="Years from Sample",
ylab="Survival")
> lines(sfit1, col=c(1,2,4), lty=2, lwd=1,
xscale=365.25)
> legend(730, .4, levels(fdata$group), lty=1, col=c(1,2,4),
bty='n', lwd=2)

5
1.0
0.8
0.6
Survival

0.4

FLC < 3.38


3.38 − 4.71
FLC > 4.71
0.2
0.0

0 2 4 6 8 10 12 14

Years from Sample

Figure 2: Survival curves from a case-control sample are shown as solid lines, dashed lines are
curves for the unweighted data set (as found in figure 1).

The survival curves for the subset data are shown in figure 2. The curve for the high risk group
is unchanged, since by definition all of those subjects were retained. We see that adjustment
for age and sex has reduced the apparent survival difference between the groups by about half,
but a clinically important effect for high FLC values remains. The curve for group 1 has moved
more than that for group 2 since the age/sex adjustment is more severe for that group.
In actual practice, case-control designs arise when matching and selection can occur before
data collection, leading to a substantial decrease in the amount of data that needs to be gathered
and a consequent cost or time savings. When a data set is already in hand it has two major
disadvantages. The first is that the approach wastes data; throwing away information in order to
achieve balance is always a bad idea. Second is that though it returns an unbiased comparison,
the result is for a very odd reference population.
One advantage of matched subsets is that standard variance calculations for the curves are
correct; the values provided by the usual Kaplan-Meier program need no further processing. We
can also use the usual statistical tests to check for differences between the curves.

> survdiff(Surv(futime, death) ~ group, data=data2)


Call:
survdiff(formula = Surv(futime, death) ~ group, data = data2)

6
N Observed Expected (O-E)^2/E (O-E)^2/V
group=FLC < 3.38 758 325 438 29.30 47.19
group=3.38 - 4.71 758 362 407 4.93 7.59
group=FLC > 4.71 758 477 319 78.39 108.68

Chisq= 113 on 2 degrees of freedom, p= <2e-16

4.2 Reweighting
A more natural way to adjust the data distribution is by weighting. Let π(a, s), a = age group,
s = sex be a target population age/sex distribution for our graph, and p(a, s, i) the observed
probability of each age/sex/group combination in the data. Both π and p sum to 1. Then if
each observation in the data set is given a case weight of

π(a, s)
wasi = (1)
p(a, s, i)

the weighted age/sex distribution for each of the groups will equal the target distribution π. An
obvious advantage of this approach is that the resulting curves represent a tangible and well
defined group.
As an example, we will first adjust our curves to match the age/sex distribution of the 2000
US population, a common reference target in epidemiology studies. The uspop2 data set is found
in later releases of the survival package in R. It is an array of counts with dimensions of age, sex,
and calendar year. We only want ages of 50 and over, and the population data set has collapsed
ages of 100 and over into a single category. We create a table tab100 of observed age/sex counts
within group for our own data, using the same upper age threshold. New weights are the values
π/p = pi.us/tab100.
> refpop <- uspop2[as.character(50:100),c("female", "male"), "2000"]
> pi.us <- refpop/sum(refpop)
> age100 <- factor(ifelse(fdata$age >100, 100, fdata$age), levels=50:100)
> tab100 <- with(fdata, table(age100, sex, group))/ nrow(fdata)
> us.wt <- rep(pi.us, 3)/ tab100 #new weights by age,sex, group
> range(us.wt)
[1] 0.7751761 Inf

There are infinite weights! This is because the US population has coverage at all ages, but our
data set does not have representatives in every age/sex/FLC group combination; there are for
instance no 95 year old males in in the data set. Let us repeat the process, collapsing the US
population from single years into the 8 age groups used previously in table 2. Merging the per
age/sex/group weights found in the 3-dimensional array us.wt into the data set as per-subject
weights uses matrix subscripts, a useful but less known feature of R.
> temp <- as.numeric(cut(50:100, c(49, 54, 59, 64, 69, 74, 79, 89, 110)+.5))
> pi.us<- tapply(refpop, list(temp[row(refpop)], col(refpop)), sum)/sum(refpop)

7
0.12
f
m
f
0.10

m
f
m
f
m
0.08

f
Fraction of population

m
f f
m m f
f f
f
0.06

m f f
m
m f
0.04

m
m m

m
0.02

f
f
m
m
0.00

50−54 55−59 60−64 65−69 70−74 75−79 80−89 90+

Age group

Figure 3: Population totals for the US reference (red) and for the observed data set (black).

> tab2 <- with(fdata, table(age2, sex, group))/ nrow(fdata)


> us.wt <- rep(pi.us, 3)/ tab2
> range(us.wt)
[1] 1.113898 34.002834
> index <- with(fdata, cbind(as.numeric(age2), as.numeric(sex),
as.numeric(group)))
> fdata$uswt <- us.wt[index]
> sfit3a <-survfit(Surv(futime, death) ~ group, data=fdata, weight=uswt)

A more common choice is to use the overall age/sex distribution of the sample itself as our
target distribution π, i.e., the empirical distribution. Since FLC data set is population based
and has excellent coverage of the county, this will not differ greatly from the US population in
this case, as is displayed in figure 3.

> tab1 <- with(fdata, table(age2, sex))/ nrow(fdata)


> matplot(1:8, cbind(pi.us, tab1), pch="fmfm", col=c(2,2,1,1),
xlab="Age group", ylab="Fraction of population",
xaxt='n')
> axis(1, 1:8, levels(fdata$age2))

8
> tab2 <- with(fdata, table(age2, sex, group))/nrow(fdata)
> tab3 <- with(fdata, table(group)) / nrow(fdata)
> rwt <- rep(tab1,3)/tab2
> fdata$rwt <- rwt[index] # add per subject weights to the data set
> sfit3 <- survfit(Surv(futime, death) ~ group, data=fdata, weight=rwt)
> temp <- rwt[,1,] #show female data
> temp <- temp %*% diag(1/apply(temp,2,min))
> round(temp, 1) #show female data
age2 [,1] [,2] [,3]
50-54 1.0 2.2 11.4
55-59 1.0 2.2 10.6
60-64 1.1 2.0 5.8
65-69 1.1 1.6 5.7
70-74 1.2 1.3 4.7
75-79 1.3 1.0 3.7
80-89 1.7 1.0 1.8
90+ 2.7 1.0 1.0

> plot(sfit3, mark.time=F, col=c(1,2,4), lty=1, lwd=2,


xscale=365.25, xlab="Years from Sample",
ylab="Survival")
> lines(sfit3a, mark.time=F, col=c(1,2,4), lty=1, lwd=1,
xscale=365.25)
> lines(sfit1, mark.time=F, col=c(1,2,4), lty=2, lwd=1,
xscale=365.25)
> legend(730, .4, levels(fdata$group), lty=1, col=c(1,2,4),
bty='n', lwd=2)
The calculation of weights is shown above, and finishes with a table of the weights for the
females. The table was scaled so as to have a minimum weight of 1 in each column for simpler
reading. We see that for the low FLC group there are larger weights for the older ages, whereas
the high FLC group requires substantial weights for the youngest ages in order to achieve balance.
The resulting survival curve is shown in figure 4. The distance between the adjusted curves is
similar to the results from subset selection, which is as expected since both approaches are
correcting for the same bias, but results are now for an overall population distribution that
matches Olmsted County. The curves estimate what the results would have looked like, had
each of the FLC groups contained the full distribution of ages.
Estimation based on reweighted data is a common theme in survey sampling. Correct stan-
dard errors for the curves are readily computed using methods from that literature, and are
available in some software packages. In R the svykm routine in the survey package handles
both this simple case and more complex sampling schemes. Tests of the curves can be done
using a weighted Cox model; the robust variance produced by coxph is identical to the standard
Horvitz-Thompsen variance estimate used in survey sampling [1]. The robust score test from
coxph corresponds to a log-rank test corrected for weighting. (In the example below the svykm
function is only run if the survey package is already loaded, as the variance calculation is very
slow for this large data set.)

9
1.0
0.8
0.6
Survival

0.4

FLC < 3.38


3.38 − 4.71
FLC > 4.71
0.2
0.0

0 2 4 6 8 10 12 14

Years from Sample

Figure 4: Survival curves for the three groups using reweighted data are shown with solid lines,
the original unweighted analysis as dashed lines. The heavier solid line adjusts to the Olmsted
population and the lighter one to the US population.

> id <- 1:nrow(fdata)


> cfit <- coxph(Surv(futime, death) ~ group + cluster(id), data=fdata,
weight=rwt)
> summary(cfit)$robscore
test df pvalue
1.076737e+02 2.000000e+00 4.158712e-24
> if (exists("svykm")) { #true if the survey package is loaded
sdes <- svydesign(id = ~0, weights=~rwt, data=fdata)
dfit <- svykm(Surv(futime, death) ~ group, design=sdes, se=TRUE)
}

Note: including the cluster term in the coxph call causes it to treat the weights as resampling
values and thus use the proper survey sampling style variance. The default without that term
would be to treat the case weights as replication counts. This same alternate variance estimate
is also called for when there are correlated observations; many users will be more familiar with
the cluster statement in that context.

10
Inverse probability weighting Notice that when using the overall population as the target
distribution π we can use Bayes rule to rewrite the weights as
1 Pr(age = a, sex = s, group = i)
=
wasi Pr(age = a, sex = s)
= Pr(group = i|age = a, sex = s)

This last is precisely the probability estimated by a logistic regression model, leading to
inverse probability weighting as an alternate label for this approach. We can reproduce the
weights calculated just above with three logistic regression models.
> options(na.action="na.exclude")
> gg <- as.numeric(fdata$group)
> lfit1 <- glm(I(gg==1) ~ factor(age2) * sex, data=fdata,
family="binomial")
> lfit2 <- glm(I(gg==2) ~ factor(age2) * sex, data=fdata,
family="binomial")
> lfit3 <- glm(I(gg==3) ~ factor(age2) * sex, data=fdata,
family="binomial")
> temp <- ifelse(gg==1, predict(lfit1, type='response'),
ifelse(gg==2, predict(lfit2, type='response'),
predict(lfit3, type='response')))
> all.equal(1/temp, fdata$rwt)
[1] TRUE

If there were only 2 groups then only a single regression model is needed since P(group 2) =
1 - P(group 1). Note the setting of na.action, which causes the predicted vector to have the
same length as the original data even when there are missing values. This simplifies merging the
derived weights with the original data set.
An advantage of the regression framework is that one can easily accommodate more variables
by using a model with additive terms and only a few selected interactions, and the model can
contain continuous as well as categorical predictors. The disadvantage is that such models are
often used without the necessary work to check their validity. For instance models with age +
sex could have been used above. This makes the assumption that the odds of being a member of
group 1 is linear in age and with the same slope for males and females; ditto for the models for
group 2 and group 3. How well does this work? Since the goal of reweighting is to standardize
the ages, a reasonable check is to compute and plot the reweighted age distribution for each flc
group.
Figure 5 shows the result. The reweighted age distribution is not perfectly balanced, i.e.,
the ‘1’, ‘2’ and ‘3’ symbols do no exactly overlay one another, but in this case the simple linear
model has done an excellent job. We emphasize that whenever the reweighting is based on a
simplified model then such a check is obligatory. It is quite common that a simple model is
not sufficient and the resulting weight adjustment is inadequate. Like a duct tape auto repair,
proceeding forward as though the underlying problem has been addressed is then most unwise.
> lfit1b <-glm(I(gg==1) ~ age + sex, data=fdata,
family="binomial")

11
3
25

+2
20

1
+
Weighted frequency (%)

3
1 3
2
+
15

1 1+
2
2 12
3 +
2 12
10

1
3 + +3
3
5

23
1+

50−54 55−59 60−64 65−69 70−74 75−79 80−89 90+

Age group

Figure 5: The re-weighted age distribution using logistic regression with continuous age, for
females, FLC groups 1–3. The target distribution is shown as a “+”. The original unadjusted
distribution is shown as dashed lines.

> lfit2b <- glm(I(gg==2) ~ age +sex, data=fdata,


family="binomial")
> lfit3b <- glm(I(gg==3) ~ age + sex, data=fdata,
family="binomial")
> # weights for each group using simple logistic
> twt <- ifelse(gg==1, 1/predict(lfit1b, type="response"),
ifelse(gg==2, 1/predict(lfit2b, type="response"),
1/predict(lfit3b, type="response")))
> tdata <- data.frame(fdata, lwt=twt)
> #grouped plot for the females
> temp <- tdata[tdata$sex=='F',]
> temp$gg <- as.numeric(temp$group)
> c1 <- with(temp[temp$gg==1,], tapply(lwt, age2, sum))
> c2 <- with(temp[temp$gg==2,], tapply(lwt, age2, sum))
> c3 <- with(temp[temp$gg==3,], tapply(lwt, age2, sum))
> xtemp <- outer(1:8, c(-.1, 0, .1), "+") #avoid overplotting
> ytemp <- 100* cbind(c1/sum(c1), c2/sum(c2), c3/sum(c3))

12
> matplot(xtemp, ytemp, col=c(1,2,4),
xlab="Age group", ylab="Weighted frequency (%)", xaxt='n')
> ztab <- table(fdata$age2)
> points(1:8, 100*ztab/sum(ztab), pch='+', cex=1.5, lty=2)
> # Add the unadjusted
> temp <- tab2[,1,]
> temp <- scale(temp, center=F, scale=colSums(temp))
> matlines(1:8, 100*temp, pch='o', col=c(1,2,4), lty=2)
> axis(1, 1:8, levels(fdata$age2))

Rescaled weights As the weights were defined in equation 1, the sum of weights for each of
the groups is 7845, the number of observations in the data set. Since the number of subjects
in group 3 is one seventh of that in group 1, the average weight in group 3 is much larger. An
alternative is to define weights in terms of the within group distribution rather than the overall
distribution, leading to the rescaled weights w∗
π(a, s)
w∗ = (2)
p(a, s|i)
P(group = i)
= (3)
P(group = i|age = a, sex = s)
Each group’s weights are rescaled by the overall prevalence of the group. In its simplest form,
the weights in each group are scaled to add up to the number of subjects in the group.
> # compute new weights
> wtscale <- table(fdata$group)/ tapply(fdata$rwt, fdata$group, sum)
> wt2 <- c(fdata$rwt * wtscale[fdata$group])
> c("rescaled cv"= sd(wt2)/mean(wt2), "rwt cv"=sd(fdata$rwt)/mean(fdata$rwt))
rescaled cv rwt cv
0.3612141 1.2772960
> cfit2a <- coxph(Surv(futime, death) ~ group + cluster(id),
data=fdata, weight= rwt)
> cfit2b <- coxph(Surv(futime, death) ~ group + cluster(id),
data=fdata, weight=wt2)
> round(c(cfit2a$rscore, cfit2b$rscore),1)
[1] 107.7 116.3
The rescaling results in weights that are much less variable across groups. This operation has
no impact on the individual survival curves or their standard errors, since within group we
have multiplied all weights by a constant. When comparing curves across groups, however, the
rescaled weights reduce the standard error of the test statistic. This results in increased power
for the robust score test, although in this particular data set the improvement is not very large.

5 Conditional method
In the marginal approach we first balance the data set and then compute results on the adjusted
data. In the conditional approach we first compute a predicted survival curve for each subject

13
that accounts for flc group, age and sex, and then take a weighted average of the curves to get an
overall estimate for each flc group. For both methods a central consideration is the population of
interest, which drives the weights. Modeling has not removed the question of who these curves
should represent, it has simply changed the order of operation between the weighting step and
the survival curves step.

5.1 Stratification
Our first approach is to subset the data into homogeneous age/sex strata, compute survival
curves within each strata, and then combine results. We will use the same age/sex combinations
as before. The interpretation of these groups is different, however. In the marginal approach it
was important to find age/sex groups for which the probability of membership within each FLC
group was constant within the strata (independent of age and sex, within strata), in this case
it is important that the survival for each FLC group is constant in each age/sex stratum. Ho-
mogeneity of membership within each stratum and homogeneity of survival within each stratum
may lead to different partitions for some data sets.
Computing curves for all the combinations is easy.
> allfit <- survfit(Surv(futime, death) ~ group +
age2 + sex, fdata)
> temp <- summary(allfit)$table
> temp[1:6, c(1,4)] #abbrev printout to fit page
records events
group=FLC < 3.38, age2=50-54, sex=F 738 28
group=FLC < 3.38, age2=50-54, sex=M 658 45
group=FLC < 3.38, age2=55-59, sex=F 638 52
group=FLC < 3.38, age2=55-59, sex=M 553 58
group=FLC < 3.38, age2=60-64, sex=F 490 52
group=FLC < 3.38, age2=60-64, sex=M 427 63

The resultant survival object has 48 curves: 8 age groups * 2 sexes * 3 FLC groups. To get
a single curve for the first FLC group we need to take a weighted average over the 16 age/sex
combinations that apply to that group, and similarly for the second and third FLC subset.
Combining the curves is a bit of a nuisance computationally because each of them is reported
on a different set of time points. A solution is to use the summary function for survfit objects
along with the times argument of that function. This feature was originally designed to allow
printout of curves at selected time points (6 months, 1 year, . . . ), but can also be used to select
a common set of time points for averaging. We will arbitrarily use 4 per year, which is sufficient
to create a visually smooth plot over the time span of interest. By default summary does not
return data for times beyond the end of a curve, i.e., when there are no subjects left at risk; the
extend argument causes a full set of times to always be reported. As seen in the printout above,
the computed curves are in sex within age within group order. The overall curve is a weighted
average chosen to match the original age/sex distribution of the population.

> xtime <- seq(0, 14, length=57)*365.25 #four points/year for 14 years
> smat <- matrix(0, nrow=57, ncol=3) # survival curves
> serr <- smat #matrix of standard errors

14
1.0
0.8
0.6
Survival

0.4
0.2
0.0

0 2 4 6 8 10 12 14

Years from sample

Figure 6: Estimated curves from a stratified model, along with those from the uncorrected fit as
dashed lines.

> pi <- with(fdata, table(age2, sex))/nrow(fdata) #overall dist


> for (i in 1:3) {
temp <- allfit[1:16 + (i-1)*16] #curves for group i
for (j in 1:16) {
stemp <- summary(temp[j], times=xtime, extend=T)
smat[,i] <- smat[,i] + pi[j]*stemp$surv
serr[,i] <- serr[,i] + pi[i]*stemp$std.err^2
}
}
> serr <- sqrt(serr)
> plot(sfit1, lty=2, col=c(1,2,4), xscale=365.25,
xlab="Years from sample", ylab="Survival")
> matlines(xtime, smat, type='l', lwd=2, col=c(1,2,4),lty=1)

Figure 6 shows the resulting averaged curves. Overlaid are the curves for the unadjusted model.
Very careful comparison of these curves with the weighted estimate shows that they have almost
identical spread, with just a tiny amount of downward shift.
There are two major disadvantages to the stratified curves. The first is that when the original

15
data set is small or the number of confounders is large, it is not always feasible to stratify into
a large enough set of groups that each will be homogeneous. The second is a technical aspect of
the standard error estimate. Since the curves are formed from disjoint sets of observations they
are independent and the variance of the weighted average is then a weighted sum of variances.
However, when a Kaplan-Meier curve drops to zero the usual standard error estimate at that
point involves 0/0 and becomes undefined, leading to the NaN (not a number) value in R. Thus
the overall standard error becomes undefined if any of the component curves falls to zero. In the
above example this happens at about the half way point of the graph. (Other software packages
carry forward the se value from the last no-zero point on the curve, but the statistical validity
of this is uncertain.)
To test for overall difference between the curves we can use a stratified test statistic, which
is a sum of the test statistics computed within each subgroup. The most common choice is the
stratified log-rank statistic which is shown below. The score test from a stratified Cox model
would give the same result.
> survdiff(Surv(futime, death) ~ group + strata(age2, sex), fdata)
Call:
survdiff(formula = Surv(futime, death) ~ group + strata(age2,
sex), data = fdata)

N Observed Expected (O-E)^2/E


group=FLC < 3.38 5560 1110 1360 46.07
group=3.38 - 4.71 1527 562 531 1.77
group=FLC > 4.71 758 477 257 187.59
(O-E)^2/V
group=FLC < 3.38 138.22
group=3.38 - 4.71 2.42
group=FLC > 4.71 230.34

Chisq= 261 on 2 degrees of freedom, p= <2e-16

5.2 Modeling
The other approach for conditional estimation is to model the risks due to the confounders.
Though we have left it till last, this is usually the first (and most often the only) approach used
by most data analysts.
Let’s start with the very simplest method: a stratified Cox model.
> cfit4a <- coxph(Surv(futime, death) ~ age + sex + strata(group),
data=fdata)
> surv4a <- survfit(cfit4a)
> plot(surv4a, col=c(1,2,4), mark.time=F, xscale=365.25,
xlab="Years post sample", ylab="Survival")

This is a very fast and easy way to produce a set of curves, but it has three problems. First
is the assumption that this simple model adequately accounts for the effects of age and sex on

16
1.0
0.8
0.6
Survival

0.4

FLC low
FLC med
FLC high
0.2
0.0

0 2 4 6 8 10 12 14

Years from Sample

Figure 7: Curves for the three groups, adjusted for age and sex via a risk model. Dotted lines
show the curves from marginal adjustment. Solid curves are for the simple risk model cfit4a.

survival. That is, it assumes that the effect of age on mortality is linear, the sex difference
is constant across all ages, and that the coefficients for both are identical for the three FLC
groups. The second problem with this approach is that it produces the predicted curve for a
single hypothetical subject of age 64.3 years and sex 0.45, the means of the covariates, under
each of the 3 FLC scenarios. However, we are interested in the adjusted survival of a cohort
of subjects in each range of FLC, and the survival of an “average” subject is not the average
survival of a cohort. The third and most serious issue is that it is not clear exactly what these
“adjusted” curves represent — exactly who is this subject with a sex of 0.45? Multiple authors
have commented on this problem, see Thomsen et al [8], Nieto and Coresh [6] or chapter 10 of
Therneau and Grambsh [7] for examples. Even worse is a Cox model that treated the FLC group
as a covariate, since that will impose an additional constraint of proportional hazards across the
3 FLC groups.
We can address this last problem problem by doing a proper average. A Cox model fit can
produce the predicted curves for any age/sex combination. The key idea is to produce a predicted
survival curve for every subject of some hypothetical population, and then take the average of
these curves. The most straightforward approach is to retrieve the predicted individual curves
for all 7845 subjects in the data set, assuming each of the three FLC strata one by one, and take
a simple average for each strata. For this particular data set that is a bit slow since it involves

17
7845 curves. However there are only 98 unique age/sex pairs in the data, it is sufficient to obtain
the 98 * 3 FLC groups unique curves and take a weighted average. We will make use of the
survexp function, which is designed for just this purpose. Start by creating a data set which
has one row for each age/sex combination along with its count. Then replicate it into 3 copies,
assigning one copy to each of the three FLC strata.
> tab4a <- with(fdata, table(age, sex))
> uage <- as.numeric(dimnames(tab4a)[[1]])
> tdata <- data.frame(age = uage[row(tab4a)],
sex = c("F","M")[col(tab4a)],
count= c(tab4a))
> tdata3 <- tdata[rep(1:nrow(tdata), 3),] #three copies
> tdata3$group <- factor(rep(1:3, each=nrow(tdata)),
labels=levels(fdata$group))
> sfit4a <- survexp(~group, data=tdata3, weight = count,
ratetable=cfit4a)
> plot(sfit4a, mark.time=F, col=c(1,2,4), lty=1, lwd=2,
xscale=365.25, xlab="Years from Sample",
ylab="Survival")
> lines(sfit3, mark.time=F, col=c(1,2,4), lty=2, lwd=1,
xscale=365.25)
> legend(730,.4, c("FLC low", "FLC med", "FLC high"), lty=1, col=c(1,2,4),
bty='n', lwd=2)
Figure 7 shows the result. Comparing this to the prior 3 adjustments shown in figures 4, 5,
and 6 we see that this result is different. Why? Part of the reason is due to the fact that
E[f (X)] 6= f (E[X]) for any non-linear operation f , so that averages of survival curves and
survival curves of averages will never be exactly the same. This may explain the small difference
between the stratified and the marginal approaches of figures 4 and 6, which were based on the
same subsets. The Cox based result is systematically higher than the stratified one, however, so
something more is indicated.
Aside: An alternate computational approach is to create the individual survival curves using
the survfit function and then take averages.
> tfit <- survfit(cfit4a, newdata=tdata, se.fit=FALSE)
> curves <- vector('list', 3)
> twt <- c(tab4a)/sum(tab4a)
> for (i in 1:3) {
temp <- tfit[i,]
curves[[i]] <- list(time=temp$time, surv= c(temp$surv %*% twt))
}
The above code is a bit sneaky. I know that the result from the survfit function contains a matrix
tfit$surv of 104 columns, one for each row in the tdata data frame, each column containing
the curves for the three strata one after the other. Sub setting tfit results in the matrix for a
single flc group. Outside of R an approach like the above may be needed, however.
So why are the modeling results so different than either reweighting or stratification? Suspi-
cion first falls on the use of a simple linear model for age and sex, so start by fitting a slightly

18
3
0.4

2
0.3

1
Deaths

crate
0.2

0
0.1

−1
−2
0.0

0 2 4 6 8 10 12 14 −2 −1 0 1 2 3

Years from sample mrate

Figure 8: Left panel: comparison of Cox model based adjustment (solid) with the curves based
on marginal adjustment (dashed). The Cox model curves without (black) and with (red) an
age*sex interaction term overlay. Right panel: plot of the predicted relative risks from a Cox
model crate versus population values from the Minnesota rate table.

19
more refined model that allows for a different slope for the two sexes, but is still linear in age.
In this particular data set an external check on the fit is also available via the Minnesota death
rate tables, which are included with the survival package as survexp.mn. This is an array that
contains daily death rates by age, sex, and calendar year.
> par(mfrow=c(1,2))
> cfit4b <- coxph(Surv(futime, death) ~ age*sex + strata(group),
fdata)
> sfit4b <- survexp(~group, data=tdata3, ratetable=cfit4b, weights=count)
> plot(sfit4b, fun='event', xscale=365.25,
xlab="Years from sample", ylab="Deaths")
> lines(sfit3, mark.time=FALSE, fun='event', xscale=365.25, lty=2)
> lines(sfit4a, fun='event', xscale=365.25, col=2)
> temp <- median(fdata$sample.yr)
> mrate <- survexp.mn[as.character(uage),, as.character(temp)]
> crate <- predict(cfit4b, newdata=tdata, reference='sample', type='lp')
> crate <- matrix(crate, ncol=2)[,2:1] # mrate has males then females, match it
> # crate contains estimated log(hazards) relative to a baseline,
> # and mrate absolute hazards, make both relative to a 70 year old
> for (i in 1:2) {
mrate[,i] <- log(mrate[,i]/ mrate[21,2])
crate[,i] <- crate[,i] - crate[21,2]
}
> matplot(mrate, crate, col=2:1, type='l')
> abline(0, 1, lty=2, col=4)
The resulting curves are shown in the left panel of figure 8 and reveal that addition of an
interaction term did not change the predictions, and that the Cox model result for the highest
risk group is distinctly different predicted survival for the highest FLC group is distinctly different
when using model based prediction. The right hand panel of the figure shows that though there
are slight differences with the Minnesota values, linearity of the age effect is very well supported.
So where exactly does the model go wrong? Since this is such a large data set we have the luxury
of looking at subsets. This would be a very large number of curves to plot — age by sex by FLC
= 48 — so an overlay of the observed and expected curves by group would be too confusing.
Instead we will summarize each of the groups according to their observed and predicted number
of events.
> obs <- with(fdata, tapply(death, list(age2, sex, group), sum))
> pred<- with(fdata, tapply(predict(cfit4b, type='expected'),
list(age2, sex, group), sum))
> excess <- matrix(obs/pred, nrow=8) #collapse 3 way array to 2
> dimnames(excess) <- list(dimnames(obs)[[1]], c("low F", "low M",
"med F", "med M",
"high F", "high M"))
> round(excess, 1)
low F low M med F med M high F high M
50-54 0.9 1.0 1.0 1.0 2.3 2.5

20
55-59 1.1 0.9 1.6 0.9 1.8 1.6
60-64 0.9 0.8 1.0 1.0 1.5 1.0
65-69 0.8 0.9 0.9 1.1 1.5 1.3
70-74 0.8 1.0 0.8 0.9 1.1 1.2
75-79 1.2 1.0 1.1 1.2 0.8 1.1
80-89 1.2 1.1 1.0 0.9 0.7 0.9
90+ 1.3 0.8 1.0 1.6 1.1 0.9
The excess risks, defined as the observed/expected number of deaths, are mostly modest ranging
from .8 to 1.2. The primary exception exception is the high FLC group for ages 50–59 which has
values of 1.6 to 2.5; the Cox model fit has greatly overestimated the survival for the age 50–54
and 55–59 groups. Since this is also the age category with the highest count in the data set,
this overestimation will have a large impact on the overall curve for high FLC subset, which is
exactly where the the deviation in figure 8 is observed to lie. There is also mild evidence for a
linear trend in age for the low FLC females, in the other direction. Altogether this suggests that
the model might need to have a different age coefficient for each of the three FLC groups.
> cfit5a <- coxph(Surv(futime, death) ~ strata(group):age +sex, fdata)
> cfit5b <- coxph(Surv(futime, death) ~ strata(group):(age +sex), fdata)
> cfit5c <- coxph(Surv(futime, death) ~ strata(group):(age *sex), fdata)
> options(show.signif.stars=FALSE) # see footnote
> anova(cfit4a, cfit5a, cfit5b, cfit5c)
Analysis of Deviance Table
Cox model: response is Surv(futime, death)
Model 1: ~ age + sex + strata(group)
Model 2: ~ strata(group):age + sex
Model 3: ~ strata(group):(age + sex)
Model 4: ~ strata(group):(age * sex)
loglik Chisq Df P(>|Chi|)
1 -15155
2 -15139 31.0215 2 1.836e-07
3 -15139 0.0615 2 0.96972
4 -15136 6.9116 3 0.07477
> temp <- coef(cfit5a)
> names(temp) <- c("sex", "ageL", "ageM", "ageH")
> round(temp,3)
sex ageL ageM ageH
0.330 0.112 0.101 0.078
The model with separate age coefficients for each FLC group gives a major improvement in
goodness of fit, but adding separate sex coefficients per group or further interactions does not
add importantly beyond that. 1
A recheck of the observed/expected values now shows a much more random pattern, though
some excess remains in the upper right corner. The updated survival curves are shown in figure
9 and now are in closer concordance with the marginal fit.
1 There are certain TV shows that make one dumber just by watching them; adding stars to the output has

the same effect on statisticians.

21
0.5
0.4
0.3
Deaths

0.2
0.1
0.0

0 2 4 6 8 10 12 14

Years from sample

Figure 9: Adjusted survival for the 3 FLC groups based on the improved Cox model fit. Dashed
lines show the predictions from the marginal model.

22
> pred5a <- with(fdata, tapply(predict(cfit5a, type='expected'),
list(age2, sex, group), sum))
> excess5a <- matrix(obs/pred5a, nrow=8,
dimnames=dimnames(excess))
> round(excess5a, 1)
low F low M med F med M high F high M
50-54 1.0 1.4 0.9 1.1 1.2 1.5
55-59 1.2 1.1 1.4 0.9 1.1 1.1
60-64 0.9 0.9 1.0 1.1 1.0 0.8
65-69 0.8 1.0 0.9 1.1 1.2 1.1
70-74 0.8 1.0 0.8 0.9 1.0 1.1
75-79 1.2 0.9 1.1 1.1 0.8 1.0
80-89 1.1 0.9 1.0 0.8 0.9 0.9
90+ 1.2 0.6 1.1 1.5 1.5 1.1
> sfit5 <- survexp(~group, data=tdata3, ratetable=cfit5a, weights=count)
> plot(sfit3, fun='event', xscale=365.25, mark.time=FALSE, lty=2, col=c(1,2,4),
xlab="Years from sample", ylab="Deaths")
> lines(sfit5, fun='event', xscale=365.25, col=c(1,2,4))

One problem with the model based estimate is that standard errors for the curves are complex.
Standard errors of the individual curves for each age/sex/FLC combination are a standard output
of the survfit function, but the collection of curves is correlated since they all depend on a common
estimate of the model’s coefficient vector β. Curves with disparate ages are anti-correlated (an
increase in the age coefficient of the model would raise one and lower the other) whereas those
for close ages are positively correlated. A proper variance for the unweighted average has been
derived by Gail and Byar [3], but this has not been implemented in any of the standard packages,
nor extended to the weighted case. A bootstrap estimate would appear to be the most feasible.

6 Conclusions
When two populations need to be adjusted and one is much larger than the other, the balanced
subset method has been popular. It is most often seen in the context of a case-control study,
with cases as the rarer group and a set of matched controls selected from the larger one. This
method has the advantage that the usual standard error estimates from a standard package are
appropriate, so no further work is required. However, in the general situation it leads to a correct
answer but for the wrong problem, i.e., not for a population in which we are interested.
The population reweighted estimate is flexible, has a readily available variance in some sta-
tistical packages (but not all), and the result is directly interpretable. It is the method we
recommend in general. The approach can be extended to a large number of balancing factors
by using a regression model to derive the weights. Exploration and checking of said model for
adequacy is an important step in this case. The biggest downside to the method arises when
there is a subset which is rare in the data sample but frequent in the adjusting population. In
this case subjects in that subset will be assigned large weights, and the resulting curves will have
high variance.

23
The stratified method is closely related to reweighting (not shown). It does not do well if the
sample size is small, however.
Risk set modeling is a very flexible method, but is also the one where it is easiest to go
wrong by using an inadequate model, and variance estimation is also difficult. To the extent
that the fitted model is relevant, it allows for interpolation and extrapolation to a reference
population with a different distribution of covariates than the one in the training data. It may
be applicable in cases such as rare subsets where population reweighting is problematic, with
the understanding that one is depending heavily on extrapolation in this case, which is always
dangerous.

7 A note on type 3 tests


One particular software package (not R) and its proponents are very fond of something called
“type 3” tests. Said tests are closely tied to a particular reference population:
ˆ For all continuous covariates in the model, the empirical distribution is used as the refer-
ence.
ˆ For all categorical adjusters, a uniform distribution over the categories is used.

Figure 10 shows the fit from such a model. Not surprisingly, the predicted death rate is
very high: 1/4 of our population is over 80 years old! The authors do not find such a prediction
particularly useful since we don’t ever expect to see a population like this (it’s sort of like planning
for the zombie apocalypse), but for those enamored of type 3 tests this shows how to create the
corresponding curves.

> # there is a spurious warning from the model below: R creates 3 unneeded
> # columns in the X matrix
> cfit6 <- coxph(Surv(futime, death) ~ strata(group):age2 + sex, fdata)
> saspop <- with(fdata, expand.grid(age2= levels(age2), sex= levels(sex),
group = levels(group)))
> sfit6 <- survexp(~group, data=saspop, ratetable=cfit6)
> plot(sfit6, fun='event', xscale=365.25, mark.time=FALSE, lty=1, col=c(1,2,4),
xlab="Years from sample", ylab="Deaths")
> lines(sfit5, fun='event', xscale=365.25, lty=2, col=c(1,2,4))

24
0.6
0.5
0.4
Deaths

0.3
0.2
0.1
0.0

0 2 4 6 8 10 12 14

Years from sample

Figure 10: Adjusted survival for the 3 FLC groups based on a fit with categorical age, and
predicting for a uniform age/sex population. Dashed lines show the predictions from the marginal
model.

25
References
[1] D. A. Binder. Fitting Cox’s proportional hazards models from survey data. Biometrika,
79:139–147, 1992.
[2] A. Dispenzieri, J. Katzmann, R. Kyle, D. Larson, T. Therneau, C. Colby, R. Clark, G. Mead,
S. Kumar, L.J. Melton III, and S.V. Rajkumar. Use of monclonal serum immunoglobulin
free light chains to predict overall survival in the general population. Mayo Clinic Proc,
87:512–523, 2012.
[3] M. H. Gail and D. P. Byar. Variance calculations for direct adjusted survival curves, with
applications to testing for no treatment effect. Biometrical J., 28:587–599, 1986.
[4] R. Kyle, T. Therneau, S.V. Rajkumar, D. Larson, M. Pleva, J. Offord, A. Dispenzieri,
J. Katzman, and L.J. Melton III. Prevalence of monoclonal gammopathy of undetermined
significance. New England J Medicine, pages 1362–1369, 2006.
[5] R. A. Kyle. “Benign” monoclonal gammopathy — after 20 to 35 years of follow-up. Mayo
Clinic Proceedings, 68:26–36, 1993.

[6] F. Javier Nieto and Josef Coresh. Adjusting survival curves for confounders: a review and a
new method. Am J of Epidemiology, pages 1059–1068, 1996.
[7] T. M. Therneau and P. M. Grambsch. Modeling Survival Data: Extending the Cox Model.
Springer-Verlag, New York, 2000.
[8] B. L. Thomsen, N. Keiding, and D. G. Altman. A note on the calculation of expected survival,
illustrated by the survival of liver transplant patients. Stat. in Medicine, 10:733–738, 1991.

26